Grid-connected battery energy storage systems require precise state-of-charge estimation to maintain reliable operation and support grid stability. Traditional estimation methods struggle to accurately model the complex electrochemical behavior of batteries, particularly under dynamic grid conditions and measurement uncertainty.
Researchers have developed a new framework called the probabilistic fractional-order Mamba Kolmogorov-Arnold network (PFO-Mam-KAN) that combines machine learning with physics-based modeling to improve SOC estimation accuracy. The system integrates three key components: a Mamba encoder that efficiently processes long sequences of battery data with minimal computational overhead, a Kolmogorov-Arnold decoder that estimates uncertain battery parameters, and a physics layer incorporating fractional-order mathematics to capture the complex, long-memory effects in battery electrochemistry.
The framework was validated using real battery data from a hybrid AC/DC microgrid test environment. Results show the system achieved 0.31% root mean square error in SOC estimation—a significant improvement over conventional approaches. The model also demonstrated robustness when exposed to realistic challenges: 3% measurement noise and temperature variations between 10°C and 40°C.
A key innovation is the adaptive Kalman-gated correction mechanism, which automatically adjusts how much the estimate relies on new measurements based on the propagated uncertainty. During uncertain conditions, the system trusts the physics model more; when measurements are reliable, it incorporates them more heavily. This balance improves accuracy while maintaining stability.
The approach is computationally efficient compared to conventional deep learning methods, making it suitable for real-time deployment in battery management systems. As utilities increasingly integrate battery storage into grids with high renewable penetration, this technology can enhance operational reliability and system-wide grid support capabilities.
